Accountability Under Surveillance: Reimagining GenAI as Learning Infrastructure for Afghan Women

The Collapse of Formal Education and the Rise of Synthetic Companionship

When the Taliban regained control of Afghanistan in 2021, the formal education system for women and girls effectively collapsed. Universities closed their doors to female students, secondary education became inaccessible, and the infrastructure of learning, mentorship, and professional socialization that typically supports educational trajectories vanished. In this vacuum, digital tools have ceased to be supplementary conveniences and have become primary lifelines for continued intellectual development. The paper Designing Safe and Accountable GenAI as a Learning Companion with Women Banned from Formal Education by Behmanush et al., accepted at ACM FAccT 2026, presents a rigorous examination of how women in Afghanistan are navigating this terrain through remote participatory design sessions with twenty participants, supported by an initial recruitment survey of 140 women.

The study reveals a profound shift in how we must conceptualize educational technology in crisis contexts. Rather than viewing generative AI as a mere information retrieval system, participants consistently framed GenAI as an always-available peer and mentor, a synthetic companionship that attempts to compensate for the absence of learning communities, career counselors, and institutional validation. This reconceptualization carries significant technical and ethical implications. The research demonstrates that accountability in such contexts is not merely about algorithmic fairness or output accuracy, but fundamentally about exposure: the digital traces left by AI interactions and the potential for those traces to trigger surveillance, household conflict, or physical danger.

The Architecture of Safety-First Design

The technical methodology employed by the research team offers a template for conducting human-centered research in high-risk environments. Using remote participatory design (PD) conducted through secure digital channels, the researchers navigated constraints including shared devices, unstable connectivity, and the necessity of concealment. Their findings coalesce into three primary themes (F1-F3) and five accountability-focused design directions (D1-D5) that merit detailed examination.

The surveillance context fundamentally inverts conventional UX optimization principles. Where commercial EdTech typically maximizes engagement metrics and session duration, the Afghan context demands minimization of detectability. Participants emphasized the need for Safety-First Interaction Design (D2), which includes features such as rapid exit mechanisms, innocuous iconography that masks educational applications as utilities or social platforms, and local data processing to reduce network traffic that might trigger scrutiny from household members or authorities. The study quantifies these concerns through qualitative coding of participant narratives regarding privacy risks and contextual mismatch, revealing that accountability must be operationalized as trace minimization rather than transparency in the traditional algorithmic sense.

Furthermore, the research identifies Contextual Mismatch (F2) as a critical failure mode. Generic GenAI systems often provide career guidance or educational pathways that assume mobility, formal credentialing, or economic resources that are simply unavailable to women under current restrictions. The design direction Aligning Localization with Realistic Opportunities (D5) therefore requires what the authors term constrained optimization: AI companions must map educational content not to global best practices, but to hyperlocal micro-economies and remote work possibilities that remain accessible under restriction. This represents a technical challenge in grounding large language models not just in language, but in geopolitical reality.

Pedagogical Integrity and the Illusion of Progress

Perhaps the most technically nuanced finding concerns the pedagogical limitations of current GenAI interfaces. Participants reported that standard direct-answer interactions create an illusion of progress, where learners feel they have mastered material because the AI has provided a coherent explanation, when in fact they have not developed independent reasoning capabilities. This risk is amplified in unsupervised self-learning contexts where no instructor provides corrective feedback or validation.

The proposed solution, Safeguarding Learning Integrity with Reasoning (D4), moves beyond simple answer provision toward Socratic questioning and structured reasoning scaffolding. The technical implementation suggested involves modifying interaction patterns to require learners to articulate their own problem-solving steps before receiving AI assistance, thereby maintaining cognitive load at pedagogically appropriate levels. This aligns with broader findings in the learning sciences regarding the importance of desirable difficulties in education, but adapts them for an environment where the AI must simultaneously serve as tutor, peer, and assessment mechanism.

The study also documents the emergence of AI-Facilitated Virtual Learning Spaces (D1), where GenAI mediates not just individual learning but collective study groups. Participants envisioned systems where the AI facilitates peer connections, moderates discussions, and maintains continuity of learning communities that physical restrictions have dispersed. This represents a shift from GenAI as tool to GenAI as infrastructure, supporting the social dimensions of learning that are often overlooked in individualistic models of computer-assisted instruction.

Participatory Design as Aspirational Intervention

The quantitative component of the study offers particularly striking evidence regarding the therapeutic and empowering potential of design participation itself. Using pre- and post-study measures, the researchers found statistically significant increases in participants' aspirations (p=.01), perceived agency (p=.01), and perceived avenues (p=.03) following their involvement in the participatory design sessions. This finding suggests that the act of envisioning technological futures serves as an intervention against the psychological toll of educational exclusion.

From a methodological perspective, this result challenges the traditional separation between research and intervention in human-subjects work. The participatory design process did not merely elicit requirements; it actively reconstructed participants' sense of possible futures. This has implications for how we conduct human-centered AI research in other marginalized contexts. Rather than treating participants as passive sources of data or requirements, the research demonstrates that collaborative speculation about AI capabilities can function as a form of cognitive restructuring, expanding the horizon of the possible even when material conditions remain constrained.

However, the study appropriately acknowledges limitations. The sample size of twenty participants, while appropriate for deep qualitative inquiry, limits generalizability across Afghanistan's diverse ethnic and linguistic communities. Additionally, the remote nature of the study, necessitated by security concerns, meant that researchers could not observe naturalistic technology use in physical spaces, potentially missing nuances of household dynamics around device sharing and surveillance. The authors also note the ethical tightrope of studying vulnerable populations without exacerbating their risk, requiring careful protocols around data storage and communication channels.

Rethinking Accountability Beyond Western Frameworks

The contribution of this work extends beyond the Afghan context to challenge foundational assumptions in AI ethics and safety research. Much of the discourse around accountable AI focuses on bias mitigation, explainability, and fairness metrics derived from Western institutional contexts. Behmanush et al. demonstrate that in gender-restrictive environments, accountability must encompass physical safety and information security as primary design constraints.

This suggests a need for context-sensitive accountability frameworks that privilege user control over data and minimization of network exposure over the transparency mandates typically emphasized in algorithmic governance. The design directions emerging from this work, particularly the emphasis on reasoning over answers and localization over universalism, offer a blueprint for GenAI deployment in other crisis contexts, including refugee education, incarceration, and regions experiencing active conflict.

The study also raises uncomfortable questions about the role of technology in entrenched political violence. While GenAI offers a valuable substitute for destroyed educational infrastructure, it risks normalizing the exclusion of women from physical educational spaces by providing a palliative rather than a solution. The researchers navigate this tension by emphasizing that accountable design must support aspirational change; the goal is not merely to help women survive under restriction, but to prepare them for professional futures should political conditions shift.

Looking forward, this research opens several avenues for technical development. Systems that can operate with minimal network signatures, that can localize career guidance to restrictive economies, and that can maintain pedagogical integrity without human oversight represent significant engineering challenges. More fundamentally, the work invites us to consider how AI systems might be designed not just for individual learners, but for communities in diaspora, maintaining the social fabric of education when physical congregation becomes impossible. The women in this study are not merely using AI to learn; they are using it to sustain intellectual community under conditions of profound isolation, suggesting that the next generation of educational AI must be designed not as tutors, but as companions in the truest sense.